LibreMarigoldV2b-depth

Marigold V2 log-stage2 adapter and fixed prompt tensors for LibreYOLO.

Usage

Requires the Marigold V2 integration and libreyolo[marigold] dependencies. The integration is on feat/marigoldv2.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreMarigoldV2b-depth.pt", device="cuda")
result = model("photo.jpg")
values = result.depth_map.numpy().data
result.plot().save("prediction.png")

Depth values use log_depth encoding and are affine-relative in that space. They are not metres. Prediction defaults to the native canvas, rounded up to multiples of 16; imgsz=512 selects a fixed square canvas. Outputs return to original resolution.

Source

Adapter and prompt tensors: huawei-bayerlab/marigold-v2-0, revision 6fd6d1ca246c9d2d99a4d8ac375a4eccc87178ad, subfolder depth/Log-stage2. Copyright 2026 Huawei Technologies Co., Ltd. Apache License 2.0. The source implementation is also Apache-2.0.

The frozen Qwen/Qwen-Image-Edit-2509 base is pinned to d3968ef930e841f4c73640fb8afa3b306a78167e and downloaded separately. Its model card declares Apache-2.0. This adapter checkpoint does not contain that base and is not a standalone model.

Modifications

LibreYOLO metadata and a variant marker are added. Inference adapter, decoder and prompt tensors are unchanged. Two named training-only iREPA projection tensors are omitted. Conversion is implemented in libreyolo/models/marigold_v2/convert.py.

Validation and limits

Two same-worker upstream/native output comparisons passed with maximum absolute difference 0.0 on NVIDIA L40S, PyTorch 2.10.0+cu128. This verifies implementation parity, not published benchmark accuracy or bitwise agreement between different machines. CUDA NF4/BF16 inference was tested. Training, export and MPS are not supported by this integration.

License

Apache License 2.0. See LICENSE and NOTICE. The pretrained weights are not covered by LibreYOLO's MIT source-code license.